From my experience working across datasets generated with different technologies, the safest approach is usually to analyze each dataset independently first, then compare the results to identify the consensus signal rather than trying to force a direct merge from the beginning.

For your case, I would separate the problem into two steps.

1. Assign cell-type identity consistently in both datasets

First, try to label the cells in each dataset using known markers for the neuronal populations you care about. You can find some helpful markers here: https://omnibusx.com/celltypes

If you have already identified a marker signature from the 10X dataset, that signature can be used as a reference when examining the FLASH-seq cells. Of course, some genes may be missing or weakly detected in the FLASH-seq dataset because of the technology, coverage, or the small number of cells, so perfect agreement is not necessarily expected. At that point, the interpretation becomes biological versus technical, and that needs to be judged carefully.

1. Identify markers for the target population independently in each dataset, then intersect them

The shared genes are usually the most convincing evidence that the same biological identity is being recovered across datasets.

With only 32 cells in the FLASH-seq dataset, I would be cautious about relying on clustering-based workflows. In this setting, a targeted validation of a known signature is often more meaningful than trying to rediscover the full structure de novo. So rather than asking whether the FLASH-seq dataset reproduces the entire clustering landscape, I would ask whether it still captures the expected marker program of the neuronal population you identified in the larger 10X dataset.

I hope this helps.

I would start with just mapping your FLASH-seq dataset onto your original 10X dataset, which is pretty straightforward with Seurat. Should let you annotate cell types and visualize them side-by-side in a clear way.